Attitude control of catamaran based on deep reinforcement learning

ObjectiveA longitudinal motion control algorithm based on deep reinforcement learning is proposed, focusing on the dependency of traditional control algorithms on precise mathematical models and system parameters in longitudinal motion control of catamarans. MethodsBy designing reward functions and neural network structures and adjusting relevant hyper-parameters, in combination with the catamaran model, through experiments, the control effect of the deep reinforcement learning DDPG algorithm and the GA-LQR algorithm under three different control modes and the robustness under different operating conditions and initial states were compared. Results Under the same operating conditions, the DDPG algorithm has a slight advantage over the GA-LQR algorithm in control effect, but its fin angle output during the control process is more aggressive. In the simulation experiments under different operating conditions and initial states, when the system and the environmental models undergo significant changes, the control effect of the DDPG algorithm is significantly affected. However, when the system and the environment undergo small changes, the DDPG algorithm exhibits better adaptability and superiority over the GA-LQR algorithm. The comprehensive analysis shows that the DDPG algorithm demonstrates similarity to the GA-LQR algorithm in terms of performance. ConclusionsThe DDPG algorithm based on deep reinforcement learning has the potential of applications in the longitudinal motion control of catamarans, providing new research directions and methodological support for ship motion control under complex sea conditions in the future.